SLIDE 1
nnsched: A neural network based process scheduler Radboud - - PowerPoint PPT Presentation
nnsched: A neural network based process scheduler Radboud - - PowerPoint PPT Presentation
nnsched: A neural network based process scheduler Radboud University Nijmegen, 12 June 2008 Peter Bex (peter.bex@xs4all.nl) Problem statement Computers are faster than a few years ago, but software is just as slow (sometimes slower)! What can
SLIDE 2
SLIDE 3
Priority
How do we know what programs are important? That depends on the user: Multimedia is important to a home user
Networking performance can be sacrificed to make multimedia apps faster.
Example: Who cares if the next episode of "The Simpsons" finishes downloading 5 minutes later? You’re watching the current episode right now and its playback must be smooth!
On a web server, networking is more important than anything else!
SLIDE 4
Priority
How do we know what programs are important? If we know what the user wants, we can look at a program’s behaviour: Does it use networking? Does it output sound? Does it process keystrokes? Does it access the harddisk? Use this information as features for a neural network which can instruct the OS scheduler
SLIDE 5
Scheduling 101
A CPU’s task: Fetch instruction Execute instruction Update instruction pointer Repeat We have only one CPU, so how come we can run more than one program at the same time?
SLIDE 6
Scheduling 101
Multitasking works like this: Run a tiny bit of program A Run a tiny bit of program B Run a tiny bit of program C Repeat Switch very quickly, and it looks like they’re running concurrently. This is a lot like how cartoons work: quickly alternate pictures to make them move.
SLIDE 7
Scheduling 101
How do we prioritize? When choosing the next program to run on the CPU, choose the program that’s important more often. This is tricky to get right! Existing OSes already allow you to define priorities. It’s a lot of work for a user to define priority this for every program he uses Let a neural network perform this task
SLIDE 8
Implementation overview
Let’s solve this the easy way: Take an existing OS
Why NetBSD?
Free/libre software -> source is available Clean design -> easy to learn and change
Let a neural net determine priority Re-use existing priority controls
UNIX "nice" values All we need to do is gather features and run the network!
Sounds easier than it is
SLIDE 9
Implementation - kernel
Kernel extensions: Feature registration
Defines features Makes it very easy to experiment with features
Scheduler advisor
Every second, recalculate priorities by running registered feature values through net
Feature monitor "device": /dev/nnfmon
For obtaining training data
Network upload: /dev/nnconf
User can upload new networks into a running kernel
SLIDE 10
Implementation - userland
Userland (normal programs): Training program
Produces networks from feature data
Testing program
To check network performance
Configuration programs
To upload networks, fetch features from the kernel And various other utilities
SLIDE 11
Features
Initial testing was done with the following features: Terminal reads, writes and read/write ratio Audio reads, writes and read/write ratio Network reads, writes and read/write ratio Disk reads, writes and read/write ratio read/write ratio: hard to calculate with network Turned out to be a useless feature
SLIDE 12
Features - ratios
read/write ratio: hard to calculate with network Turned out to be a useless feature
SLIDE 13
Features - others
Other features: very useful Reads or writes, doesn’t matter which
SLIDE 14
Features - in detail
Clusters from the previous slide make sense, but the real-time data is very chaotic:
SLIDE 15
Network topology
To smooth out the features a bit, a "memory" layer was added to the network:
Features Input & memory
2
H H
1 3 2
O "layer" network data Common Memorize previous input M1 M M I I I1
3 2
Hidden layer(s) Output layer PCB
SLIDE 16
Results
Project was a reasonable success: Network takes work out of user’s hands User-friendly: just load a pre-trained network Very easy to experiment with features OK performance overhead Room for improvement/future reasearch: Better features/more feature research
Integration with the X windowing system (HARD!) At least adding features is very easy!
Faster training algorithms (quickprop, rprop, ...) Different networks (SOM, ...)
SLIDE 17
Demo
SLIDE 18